Skip to main content

Llama Guard

Generated Evaluation Status

This section is generated from the framework manifest and compatibility report. Do not edit it manually.

  • Framework ID: llama-guard
  • Manifest: docs/external-frameworks/llama-guard.json
  • Compatibility report: ./reports/llama-guard.compatibility.json
  • Evidence class: SOURCE_REVIEWED
  • Independently reproducible: False
  • Comparative-testing claim allowed: False
  • Missing reproducibility gates: shared_test_vector, raw_output, timestamp, runtime_configuration, source_version_or_hash, replay_commands, declared_expected_outcome, independent_reproduction
  • Evaluation result: COMPATIBILITY_EVIDENCE_ONLY
  • Cycle status: FIRST_FRAMEWORK_CYCLE_COMPLETE
  • Execution authority claim: False
  • Next bounded action: Add executable observations, raw outputs, pinned versions, replay commands, and independent reproduction before making comparative-testing claims.
  • Posting source: generated compatibility report
  • Generated status is descriptive compatibility evidence only.

Generated Authored Analysis Boundary

This section is generated. Do not edit it manually.

  • Framework ID: llama-guard
  • Framework name: Llama Guard
  • Generated sections above this boundary may be rebuilt from registry, manifest, compatibility-report, and result artifacts.
  • Authored analysis below this boundary may contain interpretation, notes, and framework-specific discussion.
  • Generators must preserve authored analysis unless a future validator explicitly declares a migration path.
  • Boundary rule: generated material is descriptive compatibility evidence only and does not create certification, endorsement, adoption, proof, or operational permission.

Generated Transition Mapping

This section is generated from the framework manifest. Do not edit it manually.

FieldGenerated Value
framework_identityLlama Guard
source_referencehttps://ollama.com/library/llama-guard3
source_versiondistribution source recorded; canonical governance package not established
allowed_use_boundarysafety-category evidence only
claimsclassification of model inputs and outputs against a safety taxonomy
non_claimsno complete provenance, policy authority, standing, certification, or execution authority
input_artifact_typemodel input, output, taxonomy, prompt template, threshold, and runtime configuration
output_artifact_typesafety classification evidence
actor_or_authority_modelmodel distribution and runtime identity; authority not established
evidence_modeldistribution source plus local bounded crosswalk
policy_or_rule_modeltaxonomy and classifier behavior dependent on version and configuration
delegation_modelnot established by classifier result
decision_or_result_modelsafety category or classifier result only
execution_authority_claimfalse
receipt_or_trace_modelcaptured classifier inputs, outputs, configuration, and local compatibility record
reconstruction_modelexact model, prompt, taxonomy, threshold, runtime, and output required
SPE_overlapclassifier evidence may inform safety review
StegVerse_ecosystem_overlapexecution-boundary and semantic review evidence
fail_closed_conditionsmissing exact model version, prompt template, taxonomy, threshold, runtime, raw output, or authority overclaim

Generated mapping is compatibility evidence only.

Generated Framework Metadata

This section is generated from the external-framework registry. Do not edit it manually.

  • Framework ID: llama-guard
  • Name: Llama Guard
  • Registry status: SOURCED-CROSSWALK-PROVISIONAL
  • Testbench state: SOURCE_RECORDED_CROSSWALK_PROVISIONAL
  • Manifest path: docs/external-frameworks/llama-guard.json
  • Source reference: https://ollama.com/library/llama-guard3
  • Metadata boundary: generated metadata is descriptive only; it does not create certification, endorsement, formalism adoption, admissibility proof, or execution authority.

Evidence posture

evidence_class: SOURCE_REVIEWED
page_completeness: COMPLETE_WITH_EXTERNAL_GATES
runtime_observation: none attached
independent_reproduction: false
comparative_testing_claim_allowed: false
execution_authority_claim_allowed: false
source_limitation: distribution page is not a complete canonical model-governance record

Published scope

Llama Guard is a safety-classification model family intended to classify model inputs and outputs against a safety taxonomy.

Distribution source: https://ollama.com/library/llama-guard3

Source snapshot posture: the recorded source identifies a distributed model package, but does not provide a complete canonical record of model lineage, training, taxonomy governance, prompt contract, thresholds, or release authority.

Native terms

Model termMeaning hereStegVerse relationship
Safety taxonomyCategories used to classify content.Policy-like evidence whose version and scope must be pinned.
Classifier outputSafety label or decision text.Bounded model evidence; not an authorization result.
Prompt templateInstructions and format controlling classification.Runtime configuration that materially affects output.
Model packageDistributed weights and metadata.Source artifact requiring identity, hash, and custody evidence.

Relationship to admissibility

Llama Guard asks: How does this content classify under this model, taxonomy, and prompt?
StegVerse asks: May the related transition bind consequence under current authority and evidence?

A classifier result can contribute safety-category evidence. It does not establish policy legitimacy, complete risk coverage, actor standing, delegation, or consequence-binding permission.

Observation boundary

No public StegVerse Llama Guard runtime observation is claimed.

shared test vector: missing
raw classifier output: missing
timestamp: missing
runtime and prompt configuration: missing
model package version or hash: missing
replay commands: missing
declared expected outcome: missing
independent reproduction: missing

StegVerse analysis

CriterionCurrent result
IdentityModel package identity is incomplete without canonical release provenance and hash.
AuthorityClassifier output does not grant or revoke execution authority.
PolicyTaxonomy categories may resemble policy classes but do not inherit StegVerse legitimacy.
DelegationNot established by content classification.
EvidenceRaw labels, scores, prompt, taxonomy, and runtime data can become evidence when retained.
ReplayabilityRequires exact weights, package hash, prompt template, taxonomy, parameters, and runtime.
ReconstructabilityPartial until source provenance and complete inference trace are captured.
Failure behaviorUnknown label, parsing error, model failure, or taxonomy mismatch must fail closed.
InteroperabilityClassification can enter a Commitment Candidate as non-authorizing safety evidence.

Commit-time interoperability contract

transition_id
actor
requested_action
target_system
content_hash
model_family
model_package_reference
model_package_hash
taxonomy_reference
taxonomy_version
prompt_template_hash
runtime_parameters
raw_classifier_output
parsed_label
confidence_or_score_if_available
policy_reference
delegation_reference
execution_context
validity_window
source_timestamp

Failure classes

Failure classAppliesCurrent boundary
Taxonomy driftYesCategory meaning can change across model or taxonomy versions.
Distribution-source limitationYesPackage listing does not establish complete canonical provenance.
Semantic equivalence divergenceYesSafety label is not StegVerse ALLOW or DENY.
Threshold or prompt driftYesPrompt and parameter changes can alter classification.
Coverage gapYesUnrepresented risks cannot be treated as evaluated.
Replay divergenceYesWeights, runtime, tokenizer, or prompt changes can alter output.
Fail-open inference errorYesMissing or malformed output must not authorize execution.

Machine-readable companions

manifest: docs/external-frameworks/llama-guard.json
compatibility report: docs/external-frameworks/reports/llama-guard.compatibility.json
canonical registry: docs/external-frameworks/index.json
canonical union: static/external-frameworks/canonical-union-inventory.v1.json

Maintenance and challenge path

Maintenance owner: StegVerse-Labs/admissibility-wiki, External Frameworks audit surface.

A challenge must identify llama-guard, the model package, taxonomy, prompt, runtime, output, or source limitation at issue and provide public evidence supporting correction. A distribution listing alone cannot support stronger provenance or runtime claims.

Validation completion criteria

canonical model and taxonomy references
pinned package, weights, tokenizer, and hashes
pinned prompt template and runtime parameters
shared content vectors
predeclared expected boundaries
raw classifier outputs and parsing behavior
timestamps and runtime configuration
replay commands
independent rerun receipt
non-claim language preserved

Benchmark relevance

execution_boundary, semantic_equivalence_boundary, unknown_trajectory_boundary, evidence_freshness_boundary

Non-claims

Classifier output is not execution authority, certification, transition admissibility, or proof of complete safety coverage. The recorded distribution page is not treated as a complete canonical model-governance record.

Next safe build target

Attach one pinned model package and taxonomy fixture with content vectors, prompt hash, raw outputs, runtime configuration, expected StegVerse boundary, replay command, and independent rerun receipt.

This page reflects a bounded admissibility packet. Publication does not create standing. The reflected claim inherits only the standing reconstructable from the referenced evidence, authority, and admissibility conditions.